CIDER: Continual Interactive Distillation for Embodied Reinforcement Learning
产业动态
来源：arXiv cs.RO发布时间待核实
arXiv:2608.21899v1 Announce Type: new Abstract: Human-in-the-loop real-world reinforcement learning enables rapid acquisition of effective robotic manipulation policies for individual tasks, often within tens of minutes. Yet it remains unclear how to extend this paradigm to continual learning, where a single policy must acquire new skills without losing previously learned behaviors.